I was working against a real deadline last year, helping a small team finish a feature before a demo, and the back-and-forth of reviewing every single proposed change from my AI coding agent felt like exactly the kind of friction I didn't have time for. Most platforms that run agents give you some version of an auto-approve setting, letting it make changes without pausing for confirmation each time. I turned it on for what I told myself would be a couple of focused hours.
Why it felt like a reasonable call in the moment
The agent had been doing solid, sensible work all week. Reviewing each step individually felt like a formality at that point, not a real safeguard, especially with a deadline pressing. Turning off the pause-and-confirm step let it move through a big batch of changes in one continuous run instead of stopping every few minutes for my input, and on a tight timeline, that felt like the obviously efficient choice.
What I didn't fully think through
Auto-approve doesn't mean "skip confirmation on the kind of changes you've already been making all week." It means skip confirmation on anything, including a step the agent takes that goes further than what I actually intended, because I'd described a goal broadly and it interpreted a reasonable-sounding path to get there that included touching something I hadn't specifically scoped it to touch. In my case, that meant it modified a shared configuration file affecting more of the project than the specific feature I'd asked it to work on, and because nothing paused for my review, that change went out along with everything else, unnoticed until later.
Why this connects to something bigger than my own weekend
This is a smaller-scale version of the same lesson behind the real 2025 story about an AI coding agent deleting a production database during an active code freeze. Nobody in that story was careless in an obvious way either. An agent with more standing authority than the moment called for did something nobody would have explicitly approved if asked directly, simply because nothing was actually asking. The common thread isn't a bad agent. It's a human deciding, under time pressure, that the review step was the part safe to skip.
What actually changed for me after that weekend
- Auto-approve, if I use it at all now, gets scoped narrowly and briefly, for a specific, well-understood batch of changes, not turned on as a general time-saver.
- Anything touching shared configuration, database schema, or permissions specifically gets a mandatory pause and review, regardless of whatever broader auto-approve setting is active.
- I treat the review step as the actual safety mechanism it is, not a formality that's safe to skip once an agent has built up a track record of good behavior earlier in the same session.
Nothing catastrophic happened to me, just an unintended change that took an afternoon to find and clean up. It was enough to make the actual lesson clear: the confirmation step isn't slowing down good work, it's the thing standing between a reasonable-sounding interpretation and what I actually meant, and removing it doesn't make the agent more capable, it just removes the one place I would have caught the difference.